The original title is "The Database Was Fixed Months Ago. The Website Disagreed."

Short summary

Using Claude Code to audit a cannabis-legality database, the author discovered three failure categories: stale data, permissive defaults masking bulk-generated errors, and silent database-display mismatches. Key lesson: with multiple data copies and no reconciliation, drift isn't a risk, it's a schedule. The reproducible audit method (dump, merge, diff, verify, fix all copies together) applies to any data-heavy system.

• •AI-audited database revealed stale data from law changes, permissive defaults on unknown entries masking bulk-generation errors, and database-display mismatches from string-matching bugs
• •Root cause: maintaining four separate copies of data (fallback array, database, profile file, search index) with no reconciliation makes drift inevitable
• •Reproducible audit methodology: dump all sources, simulate the merge, diff against ground truth, web-verify contested claims, then fix all copies in one pass with comprehensive join tests

Generated with AI, which can make mistakes.